Exploratory data analysis

import os
import json
import re
import pathlib
import numpy as np
import pandas as pd
import cv2
import hvplot.pandas
import matplotlib.pyplot as plt
import tensorflow as tf

from collections import Counter
from pathlib import Path

Reading image annotations

data_path = Path('HKR_Dataset_Words_Public')
data_length = len(os.listdir(data_path.joinpath('ann')))

print(f'Number of objects in HKR_Dataset_Words_Public: {data_length}')

Number of objects in HKR_Dataset_Words_Public: 64943

def get_annotations(data_path, img_type='.jpg'):
    columns = ['width', 'height', 'gt_text', 'file_path']
    df = pd.DataFrame(columns=columns)
    anns_path = data_path.joinpath('ann')
    for ann_path in anns_path.iterdir():
        with open(ann_path, encoding='utf8') as json_file:
            ann = json.load(json_file)
        width, height = ann['size'].values()
        gt_text = ann['description']
        file_path = ann['name'] + img_type
        df.loc[len(df)] = [width, height, gt_text, file_path]

    return df

ann_df = get_annotations(data_path)

ann_df[['height', 'width']] = ann_df[['height', 'width']].apply(lambda x : [float(i) for i in x])

ann_df.sample(10)

	width	height	gt_text	file_path
14273	528.0	82.0	Атырауская	10_30_115.jpg
61399	595.0	75.0	Снуют пунцовые	9_0_583.jpg
4321	366.0	123.0	Казахстан	10_0_58.jpg
22566	226.0	89.0	Атырау	10_9_364.jpg
34276	354.0	88.0	джигит	1_31_648.jpg
26436	365.0	89.0	Каражал	12_17_134.jpg
35796	619.0	99.0	Конь догонит,	1_4_106.jpg
63705	841.0	100.0	месяц мимоезжий,	9_40_556.jpg
63329	819.0	58.0	ложится на полати	9_36_547.jpg
4976	206.0	49.0	Костанай	10_11_238.jpg

ann_df['text_length'] = ann_df.gt_text.apply(len)
ann_df['pxls_per_chars'] = round(ann_df['width'] / ann_df['text_length'])

Obtaining statistics

ann_df.describe().round(2)

	width	height	text_length	pxls_per_chars
count	64943.00	64943.00	64943.00	64943.00
mean	443.28	75.70	11.02	41.02
std	176.95	17.39	4.39	9.58
min	44.00	14.00	2.00	11.00
25%	305.00	64.00	7.00	35.00
50%	430.00	75.00	11.00	40.00
75%	565.00	87.00	14.00	45.00
max	1697.00	150.00	42.00	475.00

ann_df.describe(include=object)

	gt_text	file_path
count	64943	64943
unique	2808	64943
top	Актау	0_0_0.jpg
freq	484	1

plot_opts = dict(xlim=(0, 1200),
                 ylim=(0, 250),
                 grid=True,
                 xticks=[0, 200, 400, 600, 800, 1000, 1200],
                 yticks=[50, 100, 150, 200, 250],
                 height=400,
                 width=550
                 )

style_opts = dict(scaling_factor=0.2,
                  line_alpha=1,
                  fill_alpha=0.1
                  )

ann_df[['width', 'height']].hvplot.scatter(x='width', y='height', **plot_opts).options(**style_opts)

ann_df.hvplot.kde(y='text_length')

ann_df['gt_text'].value_counts(normalize=True)[:50].hvplot.bar(width=700,
                                                               height=450,
                                                               rot=60, line_alpha=0,
                                                               title='Label Frequencies',
                                                               ylabel='fraction of all objects')

print('Label frequencies:')
print(ann_df.gt_text.value_counts())
print(f'Average label frequency: {round(ann_df.gt_text.value_counts().mean())}')

Label frequencies:
Актау                 484
Карагандинская        476
Казахстан             474
Шымкент               473
Кокшетау              469
                     ...
свобода любить?         1
свободи любить?         1
внебесах я парить       1
в небесах я поршть      1
Вид подстоялого         1
Name: gt_text, Length: 2808, dtype: int64
Average label frequency: 23

all_chars_freq = Counter()

for descr in ann_df['gt_text']:
    all_chars_freq += Counter(descr)

all_chars = sorted(list(all_chars_freq.keys()))

chars_df = pd.DataFrame(all_chars_freq.values(),
                        columns=['frequency'],
                        index=all_chars_freq.keys())
chars_df = chars_df.sort_values(['frequency'], ascending=False)

chars_df.hvplot.bar(width=920, height=450,
                    rot=0, line_alpha=0,
                    title='Character Frequencies',
                    ylabel='frequency')

print(f'All characters:\n {all_chars}')

All characters:
 [' ', '!', '(', ')', ',', '-', '.', ':', ';', '?', 'H', 'o', 'А', 'Б', 'В', 'Г', 'Д', 'Е', 'Ж', 'З', 'И', 'Й', 'К', 'Л', 'М', 'Н', 'О', 'П', 'Р', 'С', 'Т', 'У', 'Ф', 'Х', 'Ч', 'Ш', 'Щ', 'Ы', 'Ь', 'Э', 'Ю', 'Я', 'а', 'б', 'в', 'г', 'д', 'е', 'ж', 'з', 'и', 'й', 'к', 'л', 'м', 'н', 'о', 'п', 'р', 'с', 'т', 'у', 'ф', 'х', 'ц', 'ч', 'ш', 'щ', 'ъ', 'ы', 'ь', 'э', 'ю', 'я', 'ё', 'ғ', 'Қ', 'қ', 'Ү', 'Ө', 'ө', '–', '—', '…']

def has_cyrillic(str):
    return bool(re.search('[а-яА-ЯёЁ]', str))

kazakh_chars = [c for c in all_chars if not has_cyrillic(c) and c.isalpha()]

num_kazakh_labels = 0

for label in ann_df['gt_text']:
    if any(char in label for char in kazakh_chars):
        num_kazakh_labels += 1

print(f'Kazakh characters: {kazakh_chars}')
print(f'Number of Kazakh labels: {num_kazakh_labels}')
print(f'Percentage of Kazakh labels: {round(num_kazakh_labels/data_length*100, 2)}%')

Kazakh characters: ['H', 'o', 'ғ', 'Қ', 'қ', 'Ү', 'Ө', 'ө']
Number of Kazakh labels: 240
Percentage of Kazakh labels: 0.37%

def print_imgs_with_hist(df, size=5, width=15, height=20):
    num_cols = 2
    _, ax = plt.subplots(size, num_cols, figsize=(width, height))
    i = 0
    for _, row in df.sample(size).iterrows():
        img_path = '/'.join([data_path.as_posix(), 'img', row['file_path']])
        img = cv2.imread(img_path)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        ax[i // num_cols, i % num_cols].imshow(img)
        ax[i // num_cols, i % num_cols].set_title(f'label: {row.gt_text}')
        color = ('b','g','r')
        i += 1
        for j, col in enumerate(color):
            histr = cv2.calcHist([img], [j], None, [256], [0, 256])
            ax[i // num_cols, i % num_cols].plot(histr, color = col)
            ax[i // num_cols, i % num_cols].set_title('color histogram')
            ax[i // num_cols, i % num_cols].set_xlim([0, 255])
            ax[i // num_cols, i % num_cols].set_ylim([0, 2000])
        i += 1
    plt.show()

print(f'Color format: RGB')
print_imgs_with_hist(ann_df[['file_path', 'gt_text']])

Color format: RGB

Visualizing objects with extreme feature values

extreme_df = pd.DataFrame()

for col in ann_df.select_dtypes(include=[np.number]):
    df = ann_df[[col, 'file_path']].sort_values([col]).head()
    extreme_df['min ' + col + ' file'] = df['file_path'].array
    extreme_df['min ' + col] = df[col].array
    df = ann_df[[col, 'file_path']].sort_values([col], ascending=False).head()
    extreme_df['max ' + col + ' file'] = df['file_path'].array
    extreme_df['max ' + col] = df[col].array

extreme_df

	min width file	min width	max width file	max width	min height file	min height	max height file	max height	min text_length file	min text_length	max text_length file	max text_length	min pxls_per_chars file	min pxls_per_chars	max pxls_per_chars file	max pxls_per_chars
0	11_19_49.jpg	44.0	7_42_419.jpg	1697.0	10_20_136.jpg	14.0	0_27_44.jpg	150.0	1_28_646.jpg	2	10_38_284.jpg	42	11_27_63.jpg	11.0	1_44_76.jpg	475.0
1	11_19_41.jpg	45.0	10_36_313.jpg	1489.0	1_39_222.jpg	19.0	4_26_279.jpg	149.0	1_28_645.jpg	2	10_35_307.jpg	42	9_35_575.jpg	18.0	1_44_113.jpg	475.0
2	11_16_40_.jpg	51.0	10_38_313_.jpg	1477.0	1_37_222.jpg	20.0	10_16_83_.jpg	147.0	1_28_647.jpg	2	8_36_842.jpg	23	11_14_9.jpg	18.0	1_28_76.jpg	470.0
3	11_16_41.jpg	53.0	10_35_313.jpg	1406.0	1_14_222.jpg	21.0	10_38_460.jpg	146.0	1_28_648.jpg	2	7_42_428.jpg	23	9_37_575.jpg	18.0	1_28_78.jpg	446.0
4	1_44_108.jpg	54.0	10_38_129_.jpg	1300.0	9_45_869.jpg	23.0	10_30_366.jpg	146.0	1_28_649.jpg	2	8_36_487.jpg	23	9_47_575.jpg	18.0	1_33_97.jpg	374.0

def print_extreme_imgs(df, img_col, value_col, size=5):
    for _, row in df[:size].iterrows():
        img_path = '/'.join([data_path.as_posix(), 'img', row[img_col]])
        img = cv2.imread(img_path)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        width = min(img.shape[1]/20, 10)
        height = min(img.shape[0]/20, 10)
        title = value_col + ': ' + str(row[value_col])
        plt.figure(figsize=(width, height))
        plt.imshow(img)
        plt.title(title)
        plt.show()

for i in range(0, extreme_df.shape[1], 2):
    img_col, value_col = extreme_df.iloc[:, i:i+2].columns
    print_extreme_imgs(extreme_df, img_col, value_col)